{
  "id": 99309,
  "title": "Change Giba's Trick for -1.288 LB",
  "url": "/competitions/champs-scalar-coupling/discussion/99309",
  "author_name": "",
  "post_date": "2019-07-10T10:28:29.123210800Z",
  "votes": 5,
  "comment_count": 2,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/titericz\">Giba's </a>trick in his <a href=\"https://www.kaggle.com/titericz/giba-r-data-table-simple-features-1-17-lb/comments\">kernel</a> is to use \"regression l2\" instead of \"regression l1\". I'have just tested \"huber\" instead of \"regression l2\", which performed even better.</p>\n\n<p><strong>Regression_L2;</strong>\nCV: -1.043807 \nLB: -1.176</p>\n\n<p><strong>Huber;</strong>\nCV: -1.108865 \nLB: -1.196</p>\n\n<p>Blending with <a href=\"https://www.kaggle.com/todnewman\">Tod's</a> NN <a href=\"https://www.kaggle.com/todnewman/keras-neural-net-for-champs\">kernel</a> with the same weights as Giba did (0.6 * LGB + 0.4 * NN) increases LB score from <strong>-1.281</strong> to <strong>-1.288</strong> . Changing the weight may increase the score.</p>",
  "messages": [
    {
      "id": "572003",
      "postDate": "07/10/2019 10:28:29",
      "content": "<p><a href=\"https://www.kaggle.com/titericz\">Giba's </a>trick in his <a href=\"https://www.kaggle.com/titericz/giba-r-data-table-simple-features-1-17-lb/comments\">kernel</a> is to use \"regression l2\" instead of \"regression l1\". I'have just tested \"huber\" instead of \"regression l2\", which performed even better.</p>\n\n<p><strong>Regression_L2;</strong>\nCV: -1.043807 \nLB: -1.176</p>\n\n<p><strong>Huber;</strong>\nCV: -1.108865 \nLB: -1.196</p>\n\n<p>Blending with <a href=\"https://www.kaggle.com/todnewman\">Tod's</a> NN <a href=\"https://www.kaggle.com/todnewman/keras-neural-net-for-champs\">kernel</a> with the same weights as Giba did (0.6 * LGB + 0.4 * NN) increases LB score from <strong>-1.281</strong> to <strong>-1.288</strong> . Changing the weight may increase the score.</p>",
      "rawMarkdown": "[Giba's ](https://www.kaggle.com/titericz)trick in his [kernel](https://www.kaggle.com/titericz/giba-r-data-table-simple-features-1-17-lb/comments) is to use \"regression l2\" instead of \"regression l1\". I'have just tested \"huber\" instead of \"regression l2\", which performed even better.\n\n**Regression_L2;**\nCV: -1.043807 \nLB: -1.176\n\n**Huber;**\nCV: -1.108865 \nLB: -1.196\n\nBlending with [Tod's](https://www.kaggle.com/todnewman) NN [kernel](https://www.kaggle.com/todnewman/keras-neural-net-for-champs) with the same weights as Giba did (0.6 * LGB + 0.4 * NN) increases LB score from **-1.281** to **-1.288** . Changing the weight may increase the score.",
      "votes": null
    },
    {
      "id": "573982",
      "postDate": "07/13/2019 04:31:23",
      "content": "<p>Do what works for you - but Huber apparently does not work for all feature sets.  My LB score went down slightly with Huber.</p>",
      "rawMarkdown": "Do what works for you - but Huber apparently does not work for all feature sets.  My LB score went down slightly with Huber.",
      "votes": null
    },
    {
      "id": "581345",
      "postDate": "07/21/2019 19:29:35",
      "content": "<p><a href=\"/meminozturk\">@meminozturk</a>  are you sure that later this will not have a negative effect?</p>",
      "rawMarkdown": "meminozturk  are you sure that later this will not have a negative effect?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 573982,
      "author_name": "vettejeep",
      "author_url": "",
      "post_date": "07/13/2019 04:31:23",
      "content": "<p>Do what works for you - but Huber apparently does not work for all feature sets.  My LB score went down slightly with Huber.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 581345,
      "author_name": "muhakabartay",
      "author_url": "",
      "post_date": "07/21/2019 19:29:35",
      "content": "<p><a href=\"/meminozturk\">@meminozturk</a>  are you sure that later this will not have a negative effect?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "572003": "[Giba's ](https://www.kaggle.com/titericz)trick in his [kernel](https://www.kaggle.com/titericz/giba-r-data-table-simple-features-1-17-lb/comments) is to use \"regression l2\" instead of \"regression l1\". I'have just tested \"huber\" instead of \"regression l2\", which performed even better.\n\n**Regression_L2;**\nCV: -1.043807 \nLB: -1.176\n\n**Huber;**\nCV: -1.108865 \nLB: -1.196\n\nBlending with [Tod's](https://www.kaggle.com/todnewman) NN [kernel](https://www.kaggle.com/todnewman/keras-neural-net-for-champs) with the same weights as Giba did (0.6 * LGB + 0.4 * NN) increases LB score from **-1.281** to **-1.288** . Changing the weight may increase the score.",
    "573982": "Do what works for you - but Huber apparently does not work for all feature sets.  My LB score went down slightly with Huber.",
    "581345": "meminozturk  are you sure that later this will not have a negative effect?"
  },
  "source": "meta"
}